RAG systems, a two-node GPU cluster, and a multi-agent platform — all my own builds. It started with a documentation problem at my day job testing rocket engines; I couldn't share that work, so I rebuilt the ideas from scratch on public data.
Status: Open to ML Infrastructure, AI Application, and Forward-Deployed Engineering roles · Remote (US) · Available now
Selected work
The RAG systems are public, with reproducible evals against real corpora. The cluster and agent platform are private infrastructure.
Cited, revision-aware retrieval over 321,124 real IETF RFC chunks, with zero LLM in the answer path. It resolves the real RFC 2616 → 7230–7235 six-way obsoletion exactly — by walking a supersession graph, not by hoping retrieval gets it right.
Routes a question across subject-specialized knowledge modules, binds every claim to a source span, and abstains instead of guessing — measured head-to-head against a naive RAG baseline on the identical corpus.
A rule-based checker that flags made-up claims in AI-generated text. Tested against a public benchmark, it caught fewer than 3 in 100 fabrications — and I put that number at the top of the README instead of burying it, because the failure maps exactly where this whole class of tool stops working.
A modular, offline-capable multi-agent platform — OpenAI-compatible gateway, control plane, autonomy engine, RAG, and IDE integration — with an automated adversarial peer-review step and zero hosted-LLM dependency.
A self-built two-node multi-GPU cluster serving local models with a privacy-first architecture, thermal-aware autonomy, and GPU-wedge diagnosis and recovery.
Background
I build AI infrastructure on my own hardware: a two-node GPU cluster, an agentic RAG system, a multi-agent platform. The RAG work grew out of a documentation problem I took on at my day job testing RS-25 engines, which I cannot share, so I rebuilt the ideas from scratch on public data.
By day nothing ships until it is proven, and I hold my own systems to that same standard: config is the source of truth, defaults fail closed, and every number is labeled measured or estimated instead of asserted.
Toolkit